How test data ends up in training
Language models are trained on vast amounts of internet text. Publicly available benchmarks can be freely downloaded – and, often unintentionally, end up in exactly that training data. A model that has already seen the test questions along with their correct answers can recall them rather than actually solve them.
How big the problem actually is
A study by Johns Hopkins University found signs of contamination in 29.1% of MMLU test items. Other audits found leakage rates ranging from 1% to 45% depending on the benchmark. Some studies of multilingual benchmarks measured even higher contamination rates – a sign that the problem tends to grow over time rather than shrink, since newly released benchmarks spread quickly through forums, copies, and derivative datasets.
A concrete before/after example
When contaminated test items are swapped for an unpublished, "clean" variant of equal difficulty, results often drop noticeably: in one such study, a Mistral model's score fell by up to 13 percentage points on a decontaminated GSM8K variant compared to the original result.
Why rewording doesn't reliably evade detection
Contamination isn't limited to verbatim copying. A paraphrased or translated version of a test question can also evade standard detection methods, even though it's substantively the same task as the original – and still produce an artificially inflated score.
Why this matters to you as a decision-maker
A single high benchmark score alone doesn't tell you whether a model truly has the underlying capability or already knows the test questions. Results on privately held, current, or purpose-built test questions (more on this in this course's final module) are far more meaningful than a score on a benchmark that's been publicly available for years.